NLP Capabilities - Watson, Google Or Custom AI?

Many cognitive businesses that need NLP capabilities begin their journey by trying out generic platforms such as IBM Watson, Google Vision, Amazon Rekognition etc. Although it might work just fine for a certain set of use cases, a major complaint is that it does not help solve a specific business need. This makes users complain about their experience with the performance of IBM Watson's NLP capabilities being bad. Most customers feel that they get a good start with generic platforms but they end up spending way too much time and cost to get close to a good finish.

And that can get very frustrating!

To understand why it is so difficult to get good results with these platforms, you need to understand the context in which they work. IBM has built Watson as a generic cognitive platform that can deal with all sorts of problems for all sorts of domains - Healthcare, Retail, Insurance, etc. By virtue of doing that, it has become very broad and moderately good at a lot of things but it does not outperform everyone else in one field.

We do a lot of NLP and Image recognition for Retailers and most customers that work with us have already tried Watson or Google Vision and have not achieved what they really wanted to. Because of our focus on Retail, we can extract amazingly deep insights based on NLP and Image Recognition.

When a retail customer tries to use google vision for identifying features of a dress - it tells them that this looks like a dress and some basic details. Most enterprises already know what is the nature of content that they are passing to you. They want to know something deeper than ‘this looks like a dress’. Things like does this dress have a distressed look? Can it be worn in Summer? Is it a strapless dress? Does it have ruffles?

This is what we try to extract from images:

Similarly, for NLP - the generic extraction can be very frustrating. Watson can tell you all the topics, entities and proper nouns that it spots but without a context, it is not helpful.

Let's say you want to analyze a lot of product reviews using Watson. It is likely to pick up stuff that is not deep enough and very generic. Whereas, we pull out some amazing insights from reviews to automate merchandising for Retailers. Take a look at this:

We break down product reviews into multiple areas of interest - how do customers use it, what do they use it for, what are some of the other products that they use it with, what do they like about the performance of the product.

Even within Retail, this model changes from blenders to shoes:

Imagine if we were to show activity filters for Coffee Makers? Our customers will be really disappointed.

And that's the challenge with Watson and other generic platforms. It is built as a platform that has fundamental capabilities and it lets you build custom solutions on top. If you try to use the generic capabilities, you would be disappointed.